U.S. patent application number 09/829629 was filed with the patent office on 2002-01-10 for system and method for determining neuronal morphology and effect of substances thereon.
Invention is credited to Koh, Ying Ying, Lindquist, W. Brent, Svoboda, Karel.
Application Number | 20020004632 09/829629 |
Document ID | / |
Family ID | 26891639 |
Filed Date | 2002-01-10 |
United States Patent
Application |
20020004632 |
Kind Code |
A1 |
Lindquist, W. Brent ; et
al. |
January 10, 2002 |
System and method for determining neuronal morphology and effect of
substances thereon
Abstract
A system and method provides automated detailed analysis of
microscopic neuronal cell morphology. Accordingly, the effects of
various substances on the structure and function of neurons can be
evaluated based on morphology of the neurons. An algorithm is
presented which utilizes a geometric approach for automatically
detecting and quantifying the three-dimensional structure of
dendritic spines from stacks of image data acquired using
microscopy. Results are presented on the measurement of dendritic
spine length, volume, density, and shape classification for both
static and time-lapse images of dendrites. The approaches presented
here are generalizable to other aspects of neuronal morphology.
Inventors: |
Lindquist, W. Brent; (East
Setauket, NY) ; Koh, Ying Ying; (Port Jefferson
Station, NY) ; Svoboda, Karel; (Huntington,
NY) |
Correspondence
Address: |
Jeffrey S. Steen
DILWORTH & BARRESE, LLP
333 Earle Ovington Boulevard
Uniondale
NY
11553
US
|
Family ID: |
26891639 |
Appl. No.: |
09/829629 |
Filed: |
April 10, 2001 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
60196080 |
Apr 10, 2000 |
|
|
|
Current U.S.
Class: |
600/431 |
Current CPC
Class: |
G06V 20/69 20220101 |
Class at
Publication: |
600/431 |
International
Class: |
A61B 006/00 |
Goverment Interests
[0002] This invention was funded, at least in part, under grants
from the Department of Energy, Nos. DEFG0292ER14261 and
DEFG0290ER25084. The Government may therefore have certain rights
in the invention.
Claims
1. An algorithm for determining neuronal structure by analyzing a
microscopy image, said algorithm comprising: a processing module
for processing the image and extracting neuronal structures
therefrom based on geometrical features of the neuronal structures;
and an analyzing module for analyzing the extracted neuronal
structures to determine at least one characteristic thereof.
2. The algorithm according to claim 1, wherein the image is
selected from the group consisting of static image and time-series
images.
3. The algorithm according to claim 1, wherein the processing
module performs a deconvolution process to extract the neuronal
structures.
4. The algorithm according to claim 3, wherein the extracted
neuronal structures include a plurality of dendrites which are
identified via their respective backbones.
5. The algorithm according to claim 4, wherein the processing
module detects from the plurality of dendrites a plurality of
spines as geometric protrusions relative to the backbones.
6. The algorithm according to claim 5, wherein the processing
module subjects each geometric protrusion to a protrusion criterion
to distinguish geometric protrusions associated with the plurality
of spines from geometric protrusions not associated with the
plurality of spines.
7. The algorithm according to claim 6, wherein the processing
module correlates each detached spine of the plurality of spines to
its respective dendrite of the plurality of dendrites.
8. The algorithm according to claim 5, wherein the analyzing module
analyzes each of the plurality of spines to determine the at least
one characteristic thereof.
9. The algorithm according to claim 8, wherein the at least one
characteristic thereof is selected from the group consisting of
spine length, spine density and spine volume.
10. The algorithm according to claim 9, wherein the spine length
for a spine detached from its respective dendrite is determined by
the distance from a recorded dendrite surface volume element
corresponding to the respective dendrite to a furthest spine volume
element corresponding to the detached spine.
11. The algorithm according to claim 9, wherein the spine length
for a spine fully or partially attached to its respective dendrite
is determined by the distance from the center of mass corresponding
to base boundary points associated with the fully or partially
attached spine to a furthest spine volume element corresponding to
the fully or partially attached spine.
12. The algorithm according to claim 9, wherein the spine density
is computed as the number of spines per unit length of dendritic
backbone.
13. The algorithm according to claim 9, wherein the spine volume is
computed by multiplying the ratio of maximum spine intensity to
maximum dendrite intensity by focal volume.
14. The algorithm according to claim 5, wherein the analyzing
module classifies each of the plurality of spines according to
shape.
15. The algorithm according to claim 14, wherein each of the
plurality of spines is classified in one of the following
classifications: stubby, thin and mushroom.
16. The algorithm according to claim 14, wherein the analyzing
module determines the shape of each of the plurality of spines
based on spine length, spine head diameter and spine neck
diameter.
17. The algorithm according to claim 16, wherein a spine is
classified as a thin spine if the spine length is greater than the
neck diameter; a spine is classified as a stubby spine if the neck
diameter is approximately equal to the spine length; and a spine is
classified as a mushroom spine if the spine length does not exceed
neck diameter by more than a factor of 5 and the head diameter is
greater than the neck diameter.
18. A method for determining the effect of a substance on a neuron
comprising: subjecting the neuron to the substance; imaging the
neuron to generate at least one image; subjecting the at least one
image to an algorithm which contains (i) a processing module for
processing the image and extracting neuronal structures therefrom
based on geometrical features of the neuronal structures and (ii)
an analyzing module for analyzing the extracted neuronal structures
to determine at least one characteristic thereof; and comparing the
at least one characteristic to a corresponding at least one
characteristic of a control neuron.
19. A method for determining the effect of a substance on a neuron
according to claim 18, wherein subjecting the neuron to the
substance involves entry of the substance into the neuron.
20. A method for determining the effect of a substance on a neuron
according to claim 19, wherein the entry is accomplished by a
transfection technique selected from the groups consisting of
diffusion, electroporation, viral transfer, lipid mediated
transfer, calcium phosphate precipitation, direct injection and
biollistic transfer.
21. A method for determining the effect of a substance on a neuron
according to claim 18, wherein the image is generated by laser
scanning microscopy.
22. A method for determining the effect of a substance on a neuron
according to claim 21, wherein the laser scanning microscopy is
selected from the group consisting of 2- photon exitation laser
scanning microscopy and confocal laser scanning microscopy.
23. A method for determining the effect of a substance on a neuron
according to claim 18, wherein the neuron is contained in a brain
slice.
24. A method for determining the effect of a substance on a neuron
according to claim 18, wherein the image is selected from the group
consisting of static image and time-series images.
25. A method for determining the effect of a substance on a neuron
according to claim 18, wherein the processing module performs a
deconvolution process to extract the neuronal structures.
26. A method for determining the effect of a substance on a neuron
according to claim 25, wherein the extracted neuronal structures
include a plurality of dendrites which are identified via their
respective backbones.
27. A method for determining the effect of a substance on a neuron
according to claim 26, wherein the processing module detects from
the plurality of dendrites a plurality of spines as geometric
protrusions relative to the backbones.
28. A method for determining the effect of a substance on a neuron
according to claim 27, wherein the processing module subjects each
geometric protrusion to a protrusion criterion to distinguish
geometric protrusions associated with the plurality of spines from
geometric protrusions not associated with the plurality of
spines.
29. A method for determining the effect of a substance on a neuron
according to claim 28, wherein the processing module correlates
each detached spine of the plurality of spines to its respective
dendrite of the plurality of dendrites.
30. A method for determining the effect of a substance on a neuron
according to claim 27, wherein the analyzing module analyzes each
of the plurality of spines to determine the at least one
characteristic thereof.
31. A method for determining the effect of a substance on a neuron
according to claim 30, wherein the at least one characteristic
thereof is selected from the group consisting of spine length,
spine density and spine volume.
32. A method for determining the effect of a substance on a neuron
according to claim 31, wherein the spine length for a spine
detached from its respective dendrite is determined by the distance
from a recorded dendrite surface volume element corresponding to
the respective dendrite to a furthest spine volume element
corresponding to the detached spine.
33. A method for determining the effect of a substance on a neuron
according to claim 31, wherein the spine length for a spine fully
or partially attached to its respective dendrite is determined by
the distance from the center of mass corresponding to base boundary
points associated with the fully or partially attached spine to a
furthest spine volume element corresponding to the fully or
partially attached spine.
34. A method for determining the effect of a substance on a neuron
according to claim 31, wherein the spine density is computed as the
number of spines per unit length of dendritic backbone.
35. A method for determining the effect of a substance on a neuron
according to claim 31, wherein the spine volume is computed by
multiplying the ratio of maximum spine intensity to maximum
dendrite intensity by focal volume.
36. A method for determining the effect of a substance on a neuron
according to claim 27, wherein the analyzing module classifies each
of the plurality of spines according to shape.
37. A method for determining the effect of a substance on a neuron
according to claim 36, wherein each of the plurality of spines is
classified in one of the following classifications: stubby, thin
and mushroom.
38. A method for determining the effect of a substance on a neuron
according to claim 36, wherein the analyzing module determines the
shape of each of the plurality of spines based on spine length,
spine head diameter and spine neck diameter.
39. A method for determining the effect of a substance on a neuron
according to claim 38, wherein a spine is classified as a thin
spine if the spine length is greater than the neck diameter; a
spine is classified as a stubby spine if the neck diameter is
approximately equal to the spine length; and a spine is classified
as a mushroom spine if the spine length does not exceed neck
diameter by more than a factor of 5 and the head diameter is
greater than the neck diameter.
40. A method for determining the effect of a substance on a neuron
according to claim 18, wherein the substance is selected from the
group consisting of nucleic acid, protein, peptide, carbohydrate,
lipid, metal, radiation, temperature, pH, drug, toxin, dye, virus,
vitamin and mineral.
41. A method for determining the effect of a substance on a neuron
according to claim 18, wherein subjecting the neuron to the
substance includes exposure of the neuron to at least two dyes such
that one dye illuminates the structure of a dendrite and a second
dye illuminates distribution of a target molecule in the
neuron.
42. A method for determining the effect of a substance on a neuron
according to claim 41, wherein the second dye is a fusion protein
comprising a fluorescent protein linked to a target protein of
interest.
Description
PRIORITY
[0001] This application claims priority to a U.S. Provisional
Application filed on Apr. 10, 2000 having U.S. Provisional
Application Serial No. 60/196,080; the contents of which are
incorporated herein by reference.
BACKGROUND OF THE INVENTION
[0003] 1. Technical Field
[0004] The present invention relates to the study of neurons
including neuronal development and the effects of various agents on
neurons through analysis of optical imagery.
[0005] 2. Description of Related Art
[0006] Recent large-scale genome sequencing and expression analysis
have uncovered a multitude of genes implicated in brain
development, learning and memory, regeneration, and neurological
diseases. Determining the function of these genes and other
substances necessitates advanced techniques for controlling the
timing and location of gene expression, combined with specific
assays of neuronal cell function or morphology.
[0007] Neurons are known to include dendrites, cell bodies and
axons. The areas between adjacent neurons are known as synapses. In
most synapses presynaptic axons terminate on dendritic spines.
Spines are small bulbous compartments consisting of a spine head
attached via a thin neck to the dendritic shaft. Spine necks
dfffusionally isolate spine heads from their parent dendrites,
e.g., spines restrict the diffusion of Ca.sup.2+ and other second
messengers. In addition, spines contain a variety of organelles.
Spines contain post-synaptic densities, one of the most complex
signaling assemblies, which include synaptic receptors and their
regulators as well as various structural and adhesion molecules.
Spines also contain machinery required for protein translation.
[0008] Spine growth is associated with synaptogenesis. During
periods of synaptogenesis dendrites grow filopodia, relatively long
actin rich protrusions that often make several synapses; later
these filopodia are typically replaced by mature spines. Molecular
mechanisms underlying spine genesis and stabilization are beginning
to be investigated. Since actin is highly enriched in dendritic
spines, most studies have focused on pathways associated with the
regulation of actin dynamics. Calcineurin appears to be important
in actin stabilization in spines. The Rho family of small GTPases,
including Rho, Rac, and Cdc42, regulates various aspects of the
actin cytoskeleton and also modulates dendritic structure and spine
density.
[0009] Recent experiments have revealed that important aspects of
cognitive function, such as experience-dependent plasticity, neural
integration and learning and memory are correlated with variations
in dendritic branching morphology, and with spine density and
distribution. Similarly, age-related deficits in short-term memory,
important forms of neural dysfunction have been localized, in part,
to dendrites and spines.
[0010] Recently direct measurements in mammalian brain slices have
revealed that synaptic plasticity can manifest itself in sprouting
of filopodia, and spines, in an N-methyl-D-aspartate receptor
(NMDA-R) dependent manner. On the other hand prolonged NMDA-R
activation leads to a loss of spines. Other studies have shown that
spine density and shape is controlled by background electrical and
synaptic activity. For example, brief exposure to the sodium
channel blocker TTX increases the density of dendritic spines.
Spines display subtle actin-based motility that appears to be
abolished by low levels of alpha-amino-3-hydroxy-5-met-
hyl-4-isoxalone (AMPA) or glutamate. Low levels of AMPA also block
lesion-induced spine degeneration. Thus, spines are stabilized by
low levels of activation of synaptic receptors, but grow in
response to a global reduction of activity, as well as strong focal
increase in activation of synaptic receptors. The signaling
mechanisms underlying this complex response are not understood.
[0011] These and other findings have motivated extensive efforts to
obtain quantitative descriptions of dendritic and spine
morphologies, both statically and dynamically. Due to its superior
resolution capability in revealing ultrastructures at synaptic
junctions, serial section electron microscopy (SSEM) has been used
to quantify dendritic spine structures in three-dimensions (3-D).
This is, however, a non-vital form of observation and an extremely
labor-intensive histological approach requiring the physical
sectioning of the tissue into very thin sections and detailed
manual and/or semi-automatic registration and outlining of the
structures on each serial section.
[0012] Modem fluorescence microscope methods, such as confocal
laser scanning microscopy (CLSM) and two-photon excitation laser
scanning microscopy (2PLSM) offer many advantages over SSEM, at the
expense of reduced resolution. Sectioning is achieved by limiting
the detection (CLSM) or excitation (2PLSM) of fluorescence to a
sub-femtoliter focal volume. Optical imaging is rapid and
noninvasive. The exquisite selectivity of fluorescence allows the
detection of even single molecules against a background of billions
of others. Optical microscopy thus occupies a unique niche in
biology due to its ability to perform observations in intact,
living tissue at relatively high resolution. The properties of
fluorescence microscopy images are well understood. To image
neuronal structure, neurons are labeled with fluorescent molecules
that fill the cytoplasm homogeneously. Voxel values report the
convolution of the density of fluorescent probes with the
point-spread function (PSF) of the imaging system, which is
essentially equal to the focal volume and is easily measured.
Studies of morphological plasticity based on CLSM and 2PLSM
measurements of spine length and density have been described.
[0013] Despite these advances in modem imaging techniques, the
analysis of neuronal structure has remained largely manual. The
considerable amount of time and effort required for manually
extracting spine measurements has precluded routine studies of
large amounts of data. In addition, results are not precisely
reproducible as accuracy is dependent on the skill and habituation
of the user. A few detection and estimation techniques (Rusakov et
al., Quantification of dendritic spine populations using image
analysis and a tilting dissector, J. Neurosci. Methods, 1995; 60:
11-21; Watzel, et al., Detection of Dendritic spines in
3-dimensional images, DAGM-Symposium Bielefeld, 1995; 160-167;
Herzog, et al., Restoration of three-dimensional quasi-binary
images from confocal microscopy and its application to dendritic
trees, Cogswell C J, Conchello J. and Wilson, T., editors,
Three-Dimensional Microscopy: Image Acquisition and Processing IV.
SPIE Proceedings, 1997; Kilbom et al., Delineating and tracking
hippocampal dendritic spine plasticity using neural network
analysis of two-photon microscopy, Soc. Neurosci. Abstr., 1988; 24:
422-425) of varying degrees of automation have been suggested to
overcome the tedium and improve accuracy and reproducibility of the
result, none of which has apparently been used and verified on
large data sets. Rusakov, et al., supra, applied a medial axis
construction (skeletonization) to 2-D dendritic images to obtain
spine length measurements in 2-D and estimated the corresponding
3-D measurements using a stereological sampling procedure. As the
medial axis is sensitive to surface features, manual screening of
the medial axis was required to select among spine, dendrite and
irregular surface-induced features (i.e. artifacts). Since
measurements were based solely on the medial axis, no volumetric
estimates were obtainable. Watzel, et al., supra, have also
suggested the detection of dendritic spines using medial axis based
identification. Their 3-D algorithm was restricted to images
containing a single dendrite. The dendrite backbone (`centerline`)
was extracted from the medial axis and the remaining medial axis
`spurs` branching off the backbone were used to identify candidate
spines. A length tolerance was employed to distinguish true spines
from artifact `spurs`. No further analysis beyond that for a single
dendritic image was presented. Herzog, et al., supra, employed a
3-D reconstruction technique using a parametric model of cylinders
with hemispherical ends to `fit` the shape of the dendrites and the
spines. In this method, short spines or spines with thin necks were
hard to detect and had to be manually added to the model. An
approach using neural network recognition for spines (Kilbom, et
al., supra) has also been suggested.
[0014] The need to understand the structure and function of neurons
is a continuing one. Similarly, there is a continuing need to
determine the effects of various substances on the development,
structure and function of neurons.
SUMMARY OF THE INVENTION
[0015] An algorithm for determining neuronal structure by analyzing
a microscopy image is provided wherein the algorithm includes a
processing module for processing the image and extracting neuronal
structures therefrom based on geometrical features of the neuronal
structures, and an analyzing module for analyzing the extracted
neuronal structures to determine at least one characteristic
thereof.
[0016] Also provided is a method for determining the effect of a
substance on a neuron which includes subjecting the neuron to the
substance, imaging the neuron to generate at least one image,
subjecting the at least one image to an algorithm which contains
(i) a processing module for processing the image and extracting
neuronal structures therefrom based on geometrical features of the
neuronal structures and (ii) an analyzing module for analyzing the
extracted neuronal structures to determine at least one
characteristic thereof, and comparing the at least one
characteristic to a corresponding at least one characteristic of a
control neuron. In addition, dual-color images can be created
wherein one color is used to determine the structure of dendrites
and spines and another color is used to measure the distribution of
various cellular proteins.
BRIEF DESCRIPTION OF THE DRAWINGS
[0017] FIGS. 1 (a)-(c) illustrate a raw microscopy image, and
deblurred microscopy images after 5 and 20 iterations;
[0018] FIGS. 2(a)-(b) illustrate a medial axis from a segmented
image of a dendritic image with arrows indicating loop (L) formed
by overlapping spines, separate skeletons for disconnected features
(D) and spurious cell debris, and two backbones extracted from this
medial axis;
[0019] FIGS. 3(a)-(c) are diagrams illustrating a projected view of
12 spine candidates (shaded gray) with d.sub.b(S)=12, a sketch
indicating distances and spine candidates used in the spine
detection algorithm of the present invention; and an ideal spine
candidate symmetric along the line SE;
[0020] FIG. 4 is an illustration in 2-D of the orientation
criterion used to determine if detached (D) or attached (A) spine
components need to be merged;
[0021] FIGS. 5(a)-(b) provide a comparison of manual and automatic
spine detection on a segment of a hippocampal CA1 dendrite,
respectively;
[0022] FIG. 6 provides a comparison of manual and automatic
measurements of individual spine length (left), average spine
length (center) and spine density (right) for the dendrite shown by
FIGS. 5(a)-(b);
[0023] FIG. 7 illustrates charts of spine volume-length
scatterplots according to determined spine type for all spines in
experiments E.sub.1 and E.sub.2;
[0024] FIGS. 8(a)-(b) illustrate charts showing the number of
spines detected in the image at each time step and the detection
history of 52 spines followed for 25 minutes, e.g., 1 spine was
detected in each image taken over the 25-minute period, where spine
14 was seen sporadically over the entire period;
[0025] FIGS. 9(a)-(b) illustrate charts showing measured
distributions of spine length and volume as a function of time for
the population of 52 spines followed in a time-series of images,
and measured distribution of spine motility index fitted to an
exponential decay function; and
[0026] FIG. 10 illustrates charts showing lengths of five spines
plotted as a function of time showing comparison between manual
(close circles) and automated (open circles) measurements.
[0027] FIG. 11(A) depicts three deblurred 2PLSM images of neurons
labeled with GFP. Left image shows neurons cotransfected with
wild-type mTOR kinase; center image shows GFP transfected control
neurons; right image shows neurons cotransfected with an inactive
mTOR kinase mutant.
[0028] FIG. 11(B) graphically illustrates the results of analysis
using an algorithm according to the present invention which allows
comparison of spine density of GFP transfected neurons, mTOR
transfected neurons and inactive mTOR transfected neurons.
[0029] FIG. 12(A) depicts three deblurred 2PLSM images of neurons
labeled with GFP. Left image shows neurons cotransfected with
neuroligin (NLG); center image shows GFP transfected control
neurons; right image shows neurons cotransfected with an NLG mutant
designated AChE.
[0030] FIG. 12(B) graphically illustrates the results of analysis
using an algorithm according to the present invention which allows
comparison of spine density and spine length of GFP transfected
neurons, NLG transfected neurons and AChE transfected neurons.
DESCRIPTION OF PREFERRED EMBODIMENTS
[0031] The present invention provides efficient, detailed automated
analysis of axonic, dendritic and spine morphologies, thus allowing
assessment of the effects of various genes and other substances on
neurons. Spines participate in cell-to-cell contact and their
growth and retraction is powered by actin based motility. The
signaling networks underlying spine responses are controlled by
neurotransmitter receptors and neural activity. Naturally and
artificially induced perturbations in the signaling networks result
in alterations in axonic, dendritic and spine morphologies and
dynamics which can be measured according to the present invention
with speed and a degree of accuracy and consistency previously
unknown.
[0032] An automatic dendritic spine detection and analysis
algorithm appropriate for 3-D images obtained via laser scanning
microscopy or other type of microscopy is presented. The algorithm
of the present invention uses a geometric approach; it is highly
automatic and contains only a few parameter settings. It can be
applied to static images as well as time-series images. There is no
limitation on the number or the structure of the axons or dendrites
in the image. In addition to spine length and density, volumetric
measurements and spine classifications are obtained using the
algorithm. Finally, a simple extension of the algorithm allows the
measurement of distribution of proteins in dendrites and
spines.
[0033] The automatic dendritic spine detection and analysis
algorithm of the present invention offers an objective and
consistent analysis, requiring minimal amount of supervision and
makes accessible 3-D morphological characterizations of spine
length, volume, shape classification and spine density. Comparison
of results on spine length and density between the manual and
automatic approach of the present invention on a large number of
samples have validated the automatic approach for both static and
time-lapse, i.e., time-series, images.
[0034] The automated analysis greatly enhances speed, consistency
and objectivity. The timing results provided below show that
automated analysis of time-lapse data consisting of 50 images
tracking a total of 30-50 spines takes about 4 hours CPU time on a
Pentium.TM. III 500 Hz processor, whereas for manual analysis, an
experienced user will typically take 12 to 16 hours.
[0035] For static images the time savings compared to manual
analysis may not be as significant if experimental conditions vary
so significantly that new parameters have to be determined for each
image to be analyzed. However, automated analysis using the
algorithm of the present invention still provides more detailed,
complete and objective quantification than manual analysis.
[0036] The automatic algorithm of the present invention assumes
that the spines are simply connected to the dendrites. Small
looping structures in the medial axis indicate pairs of spines that
are too close to be resolved by imaging or segmentation. If it
satisfies the protrusion criterion each such structures will be
detected as a single entity. Resolution of such structures as
paired spines is currently not implemented. These occurrences are
estimated to affect less than 2% of the spine population.
[0037] The inventive algorithm enables calculation of spine volume,
previously not possible with manual analysis. Volumetric
measurements offer insight into understanding the electrical
capacity of the spines and the structural and electrophysiological
properties of neuronal dendrites. It is therefore an important
parameter for characterizing dendritic spines. Spine volume has not
been reported previously in any of the prior art automatic methods.
The volume measurements agree reasonably well with an SSEM analysis
of similarly aged animals, even though the preparation techniques
for the specimens are entirely different.
[0038] The various morphology-based measurement capabilities
presented below allows application of the investigation of the
functional significance of dendritic spines and their plasticity to
a wide spectrum of experimental and pathological conditions.
Automatic morphometry significantly improves the scale and accuracy
of such studies. Although not absolutely essential to the present
invention, in certain embodiments, it may be assumed that the image
to be analyzed is of a bi-phase medium, with one phase being the
neuronal cytoplasm (dendritic phase), the other being the
background tissue.
[0039] In accordance with the present invention, a substance is
contacted with and/or inserted into a neuron and the effects of the
substance on morphology is evaluated based on the observations
relating to spine length and density, volumetric measurements and
spine classifications according to the present algorithm. After a
neuron is analyzed by the present algorithm, it may be compared to
a control neuron to determine the differences, if any, that are
associated with administration of a substance of interest to the
neuron. A control neuron includes any neuron used for comparison
purposes. For example, a neuron which has been subjected to a
substance of interest and an imaging dye can be compared to (i) a
neuron containing that imaging dye or another imaging dye only,
(ii) a neuron containing that imaging dye or another imaging dye,
and another substance of interest, or (iii) a neuron containing
that imaging dye or another imaging dye, and a different
concentration of the substance of interest.
[0040] Substances which can be investigated include any known
material which is amenable to be applied to a neuron. Such
substances include, but are not limited to, chemical and/or
physical agents such as microorganisms, DNA, RNA, proteins,
peptides, carbohydrates, lipids, drugs, radiation, temperature, pH,
and diagnostic agents. For example, genes, growth factors, enzymes,
hormones, metals, viruses, bacteria, toxins, dyes,
electromagnetism, gamma radiation, neurotransmitters, electrolytes,
vitamins, minerals, antibiotics, anesthetics, antivirals,
antiseptics, antimicrobials, antiinflammatories, analgesics,
steroids, calcium channel blockers, antiarrhythmics, psychotropics,
antidepressants and the like may all be contacted with or inserted
into neurons for evaluation. The substance may be known, or not
known, to have physiological effects since one aim of the present
invention is to determine whether or not a substance has any effect
on a neuron.
[0041] Methods for subjecting neurons to a substance (including
causing a substance to enter into neuronal cells) are well known.
For example, passive or active diffusion utilizing concentration
gradients (osmolarity), solubilizers, permeation enhancers such as
aprotic solvents, e.g., DMSO, may be utilized to effect entry.
[0042] In a preferred embodiment, genetic engineering techniques
are utilized to impart substances into neurons. "Transformation" or
"transfection" (used interchangeably herein) of neuronal cells
refers to delivery of nucleic acid (DNA or RNA)) into a neuron by
any method. Suitable expression vectors include, but are not
limited to plasmids, cosmids, phage, phagemids, artificial
chromosomes and the like. Transfection of host cells can be
accomplished by, e.g., electroporation, viral transfer, lipid
mediated transfer, calcium phosphate precipitation, direct
injection and biollistic transfer to name a few.
[0043] For example, electroporation is suitable for introducing
macromolecules, including, but not limited to, DNA, RNA, dyes,
proteins and other various chemical agents, into neuronal cells.
Electroporation refers to the permeabilization of cell membranes by
application of short duration electric field pulses, traditionally
between relatively large plate electrodes. During the electric
pulse, charged macromolecules, including DNA, are actively
transported by electrophoresis across the cell membrane through
these pores. Noncharged molecules can also enter the pores by
passive diffusion. Upon pulse termination, pores reseal over
hundreds of milliseconds as measured by recovery of normal membrane
conductance values.
[0044] Biollistic transfer is a preferred method for transfection
herein and refers to any method for introducing foreign molecules
into a cell using velocity driven microprojectiles such as tungsten
or gold particles. Such velocity driven methods typically originate
from pressure bursts which include, but are not limited to, helium
driven, air driven, and gunpowder driven techniques. In biollistic
gene transfer, a desired substance, e.g., a plasmid containing a
nucleic acid sequence of interest, is precipitated on polymeric or
metallic beads. Indeed, combinations of genes can be delivered by
precipitating multiple plasmids onto beads. Thus, for example,
neurons can be transfected with DNA encoding red fluorescent
protein to determine cell structure, a chimera between a gene of
interest and DNA encoding green fluorescent protein to track
distribution of the gene product and a third plasmid of a specified
function. Biollistic gene transfer also allows accurate control
over the number of plasmids which may be introduced into the target
cell which allows a measure of control over the amount of target
gene products in the cell.
[0045] Thus, oliogonucleotides, chimeric genes, fusion proteins,
ligands, receptors, molecular labeling systems such as fluorescent
molecules, radiolabels, antibodies, antigens, avidin, streptavidin,
biocytin, and biotin are examples of substances which are suitable
for transfection herein. Examples of fluorescent molecules
(fluorochromes), include green fluorescence protein (GFP), color
shifted mutants of GFP including red shifted mutants, yellow
shifted mutants and blue shifted mutants, amino coumarin acetic
acid (AMCA), fluroscein isothiocyanate (FITC), tetramethylchodamine
isothiocyanate (TRITC), Texas Red, Cy3.0, Cy5.0 and dextran
conjugates of fluorochromes. Such labels may be used independently
or coupled to other molecules such as antibodies, antigens, avidin,
streptavidin, and nucleic probes. It is contemplated that DNA or
RNA encoding fluorescent proteins or other labeling moieties may be
fused to DNA or RNA encoding a desired protein prior to
transfection of such a chimeric molecules into the neuronal
cell.
[0046] The ability to transfer multiple genes into cells is helpful
in studies of the interaction between different proteins.
Transferring genes for colored proteins, such as green fluorescent
protein, has proven immensely useful for labeling cells in order to
visualize their shapes. These dyes can also be attached to or
otherwise coadministered with other proteins in order to see where
these proteins are located and move within the cell. As
demonstrated below, transfection was used to fill brain cells with
protein dyes in order to observe cell morphology. By introducing
other genes along with the gene for a fluorescent protein, the
effects of various proteins on neuronal cell growth can be
observed.
[0047] In a preferred embodiment, the neurons being evaluated are
studied in preparations that are as intact as possible.
Accordingly, while individual isolated neurons can be evaluated
using the methods described herein, a living brain slice offers an
attractive compromise between the limitations of cultured neurons
(e.g., limited synaptic plasticity) and the experimental
difficulties of working with intact animals. For example, brain
slices may be cultured on membranes according to procedures
described in Maletic-Savatic et al., Science, 283, pp. 1923-1927
(1999) and Stoppinin et al., J. Neurosci. Methods, 37, pp. 173-182
(1991), the contents of each being incorporated herein by
reference. Neurons in cultured brain slices preserve many of the
aspects of neural functions including normal membrane properties,
spine morphologies and robust synaptic plasticity. It should be
understood, however, that the methods of the present invention are
also applicable to use in primary cell cultures.
[0048] Any suitable technique involving microscopy to adequately
image neurons known to those skilled in the art may be utilized in
accordance with the present invention. Serial section election
microscopy, confocal laser scanning microscopy (CLSM) and
two-photon excitation laser scanning microscopy (2PLSM) may be
used. A preferred embodiment involves 2PLSM since it allows imaging
neurons in intact neural tissues. 2PLSM is well-suited to image
fluorescent molecules. Sectioning may be achieved by limiting the
exitation of fluorescence to a subfemtoliter volume. 2PLSM allows
imaging of small fluorescent molecules, e.g., GFP, filled
structures such as dendritic spines and axon terminals, even in
intact brain nervous tissue. 2PLSM also allows detection of
submicron spine length and density changes involved in synaptic
plasticity in brain slices and the intact brain. Furthermore, 2PLSM
allows measurement of movement of molecules with submicron
resolution. 2PLSM also allows the excitation of two fluorophores
with different emission wavelengths with the same excitation
wavelength, facilitating the simultaneous measurement of neuronal
structure (in one fluorescence channel) and the distribution of
selected proteins (detected through a second fluorescence
channel).
[0049] The ability to assess the effects of substances on neurons
according to the present invention provides an efficient modality
to determine what gene products or combinations of gene products
are involved in the structure of axons and/or dendrites. Indeed,
structure/function relationships may be uncovered and evaluated
according to the present invention. For example, mutants of
naturally occurring proteins can be evaluated for their effects on
neurons, e.g., dominant negative and constitutively active forms of
enzymes. Genes that control particular functions in neurons can be
identified and studied by transfecting with libraries of transgenes
and breaking the libraries into smaller pools. Moreover, the
effects of pharmacologic agents on neurons can evaluated according
to the present invention.
[0050] The description below describes the algorithm used to
analyze 3-D scanning microscopy images of fluorescent neuronal
structures (Section A). The images collected are analyzed to
include, but are not limited to, detection and measurement of
dendrites; dendritic spines; synapses; distributions of subcellular
components using chimeric proteins; and synaptic function using
genetically encoded fimctional probes. The algorithm includes a
processing module and an analyzing module for performing the
following steps.
[0051] An image is first processed by deconvolution and the
dendritic phase extracted (Section A.1). The dendrites are
identified via their backbones (Section A.2), which are extracted
from a medial axis construction. Spines are tiny appendages
attached to dendrites, and because of their small size, they can
often only contain a few dye molecules and thus show only dim
fluorescence. They are adjacent to larger, brighter dendrites and
they therefore are to detected against a hazy background. Unlike
the prior art, the spines are not detected from the medial axis
branches emerging from the backbone as it is difficult to
distinguish true spines from artifacts by this procedure. Instead,
according to the present invention, the spines are detected as
geometric protrusions relative to the backbone (Section A.3).
[0052] Each protrusion is subjected to a protrusion criterion to
distinguish true protrusion from non-spine dendrite surface
irregularity. As very thin necks are too weak to be detectable,
some spine "heads" appear to be separated from the dendrite and are
detected as detached components. After initial detection, a search
is implemented to associate detached components with their
appropriate bases. For time-series data, in which the same
dendritic branch is imaged over a sequence of time intervals,
translational effects in time are corrected for and individual
spines are then traced (Section A.4) through the time ordered
sequence of images. Finally, morphological characterizations of the
population of detected spines are extracted (Section A.5).
[0053] The steps described in Sections A.1-A.4 are preferably
performed by the processing module, while the steps described in
Section A.5 are preferably performed by the analyzing module of the
automatic dendritic spine detection and analysis algorithm
[0054] An imaging setup and the biological preparation used to
obtain data for testing and verification of these algorithms are
summarized in Section B and the Examples herein. Results of the
application of these algorithms to the analysis of hippocampal CA1
neurons and a small number of hippocampal CA3 neurons are presented
in the Examples as well.
[0055] A. Image Analysis
[0056] A.1 Image Deconvolution and Segmentation
[0057] The intrinsic spatial resolution limits of optical
microscopy arise from the diffraction of light; light from a point
source is ideally imaged to a larger spot characterized by the Airy
finction. The measured spread resulting from a given optical setup
is referred to as the point-spread-function (PSF). As a result, the
intensity recorded in any voxel (volume element) of a digitized
image is a convolution of intensities from its neighborhood.
[0058] Deconvolution is used to correct aspects of the image
degradation due to the PSF. A variety of deconvolution techniques
are available which employ either theoretical or experimental
measures of the PSF. In addition, blind deconvolution methods can
be employed which, concurrently with the deconvolution, reconstruct
an estimated PSF of the image. However, the presence of noise and
the band-limited nature of the PSF limit the improvements by means
of classical deconvolution techniques. Therefore some blurring will
remain even after deconvolution due to a trade-off between
sharpening of the image and noise amplification.
[0059] In addition, the photomultiplier tube (PMT) detectors used
in most laser scanning microscopes are "noisy". Even in darkness
PMTs produce spontaneous bright pixels which may be referred to as
"shot noise". One can deal with shot noise by applying a median
filter to the image. Median filters are known in the art. See,
e.g., Tukey, Exploratory Data Analysis, Addison-Wesley, Reading, MA
(1971). For example, the median filter may be a non-linear, lowpass
filter which replaces the greyscale value of each voxel v in the
digitized image by the median greyscale value of v and its 26
neighbors. This effectively removes shot noise but not real spines
which, under typical magnifications employed, have an effective
width covering many voxels.
[0060] A prior art iterative reblurring deconvolution algorithm was
applied to the median filtered image, which requires either a
theoretical or experimentally measured PSF. See, Kawata and
Ichioka, J. Opt. Soc. Am., 1980; 70:762-772, the contents of which
are incorporated herein by reference. Briefly, iterative reblurring
proceeds as follows. Let o.sup.(0) (x, y, z) denote the
experimental 3-D image and h (x, y, z) an appropriate PSF. Let *
and .sunburst. denote the convolution and correlation operators.
The deconvolved image .sup.(k) (x, y, z) in the k-th iteration
is
.sup.(k)=.sup.(k-1)+{o.sup.(0).sunburst.h-.sup.(k-1)*(h.sunburst.h)}.
[0061] A non-negativity constraint is applied to .sup.(k) at the
end of each iteration.
[0062] For the images analyzed in the Examples herein, the PSF was
measured by imaging a number of sub-resolution microspheres and
averaging their individual PSFs to reduce noise. FIG. 1
demonstrates the result of iterative reblurring a raw image (FIG.
1(a)) after 5 (FIG. 1(b)) and 20 (FIG. 1(c)) iterations.
Preferably, one employs k.sub.max=5 iterations of deblurring. It
should be understood that those skilled in the art may use other
deconvolution algorithms for reblurring.
[0063] Segmentation is a generic imaging term for labeling each
voxel in a greyscale (or color) image with an integer identifier
designating its "population type". For dendritic morphometry, this
requires distinguishing neuron voxels from the background tissue
voxels. A large number of segmentation algorithms are available.
See, e.g., Pal and Pal, Pattern Recognition, 1993; 26; 1277-1294.
As the dendritic images are processed first by median filtering and
deconvolution, simple thresholding is used for this final
segmentation step; all voxels of intensity greater than a threshold
value are identified as neuron, otherwise as background. In
general, a trade-off between selection of dim spines and reduction
of noise on the dendrite surface is made in selecting the
threshold.
[0064] A.2 Dendritic Backbone Extraction
[0065] Geometric analysis of a 3-D irregularly shaped object is
difficult; such analyses typically employ models based upon
geometrically simple "unit" objects. The algorithm of the present
invention builds upon the medial axis algorithm disclosed by Lee et
al., "Building skeleton models via 3-D medial surface/axis thinning
algorithms," CVGIP: Graph. Models Image Process., (1994); vol. 56:
462-478, to provide a skeleton from which the backbone of each
dendrite in an image can be extracted.
[0066] Intuitively, the medial axis captures a geometrically
faithful skeleton (consisting of curve segments joining at
vertices) of an object. In a digitized image these curve segments
consist of linked sequences of voxels, with the vertices being
voxels at which these segments join together. An example of the
medial axis of a portion of a segmented dendritic image is shown in
FIG. 2(a). (This is a view perpendicular to the optical axis.) The
medial axis obtained for the dendritic phase contains the backbone
("centerline") of each dendrite as a subset.
[0067] In addition to the backbone however, the medial axis
contains "spurs" and other features which correspond to
spine-related or non-spine-related surface features (e.g.,
incipient dendritic branches); to surface artifacts resulting from
digitization effects, segmentation errors, and boundary effects due
to the finite imaged volume; or to spurious cell "debris". Due to
resolution limits, spines emerging near each other may appear to
have overlapping tips in the digitized image, resulting in the
appearance of small loops in the medial axis (top of the parent
branch in FIG. 2(a)). A separate skeleton for each disconnected
component of dendritic phase is also contained in the medial
axis.
[0068] From the medial axis, the backbone for each dendrite is
extracted (FIG. 2(b)) in two steps. The first step is achieved by
removing the medial axis segments corresponding to all disconnected
dendritic components, and trimming short spurs and loops on the
medial axis. Removal of long "spurs" is problematic as they may
correspond to filopodia or incipient branches of the dendrite.
These long "spurs" are dealt with in the next step.
[0069] In the second step, a backbone is traced through each
dendritic branch employing a decision based upon minimum deviation
angle whenever a vertex on the trimmed skeleton is encountered. If
necessary, the number, n, of dendrites in the image can be
specified so that only the n longest backbones are retained. Any
remaining medial axis segment that is not part of a traced backbone
removed. The final set of dendritic backbones extracted from FIG.
2(a) is shown in FIG. 2(b).
[0070] A.3 Spine Detection
[0071] With the dendritic backbones isolated, the spine detection
algorithm of the present invention proceeds in four steps:
detection of detached dendritic phase components (Section A.3.a);
detection of attached spine components (Section A.3.b); elimination
of spurious or incomplete spine components (Section A.3.c); and
merging of spine components (Section A.3.d). Since a spine may be
composed of one or more detached pieces and possibly an attached
base in the segmented image, the identification of any spine is not
finalized until all four steps have been completed.
[0072] A.3.a Detached spine component detection
[0073] Dendritic phase components disconnected from the
backbone-containing dendrites are detected and tentatively
identified as detached spine components. For each detached spine
component a record is kept of its center of mass, the closest
dendritic backbone voxel and the dendritic surface voxel lying on
the line joining the center of mass to the backbone voxel. Detached
dendritic phase components that are further from the nearest
dendrite surface voxel than a maximum distance tolerance are
interpreted as false positive signals and ignored. For the images
analyzed in the Examples herein, the length tolerance is 6 .mu.m.
It should be understood that the length tolerance can be adjusted
when appropriate.
[0074] A.3.b Attached spine component detection
[0075] Ignoring the detached spine components, every dendrite phase
voxel v is labeled with a distance, d.sub.b(v), to its closest
backbone voxel. Thus tips of protrusions on the dendrite surface
are assigned the largest distances. These tip voxel locations are
then processed in descending order of d.sub.b.
[0076] For each tip voxel S, a sequence {C.sub.i}, i=1, . . .
,d.sub.b(S), of candidate spines is generated. Candidate C.sub.i
consists of all voxels w whose distance d.sub.S (w) from S is
d.sub.S(w). FIG. 3(a) shows a 2-D projected view of the candidates
(gray voxels) C.sub.1, . . . ,C.sub.12 for a tip S having
d.sub.b=12. The smaller candidates clearly contain insufficient
voxels to correctly represent the spine, whereas the larger
candidates protrude too far below the dendrite surface. The optimal
choice of a spine candidate would terminate at the surface of the
dendrite. This is achieved by estimating for the local thickness of
the dendrite as explained below.
[0077] To estimate the local thickness and choose the optimal
candidate, a ring of "spine-surface boundary points" are determined
for each candidate. For clarity of explanation, the algorithm is
first illustrated in 2-D assuming the image is projected onto its
focal plane. The 3-D algorithm is described afterwards. Two
"surface boundary voxels" P.sub.1 and P.sub.2 on the 2-D projection
are shown in FIG. 3(b). For each candidate C.sub.i, along with the
two surface boundary voxels P.sub.1.sup.C.sub.i and
P.sub.2.sup.C.sub.i, a "base voxel" E.sup.C.sub.i having the
furthest penetration into the dendrite is also determined. An
idealized sketch of such a projection illustrating S, P.sub.1,
P.sub.2 and E is shown in FIG. 3(b). These surface points are used
to determine the best measure of the dendrite thickness as follows.
A "reference candidate" C.sub.R is selected to be the smallest
volume candidate with the minimum surface to backbone distance,
d.sub.p.ident.min.sub.i{d.sub.b- (P.sub.1.sup.C.sub.i),
d.sub.b(P.sub.2.sup.C.sub.i)}. A reference candidate C.sub.R and
the distance d.sub.P are illustrated in FIG. 3(b).
[0078] In an ideal case 2 d.sub.p is the width of the dendrite and
the best candidate for the spine would be C.sub.i where
j=d.sub.b(S)-d.sub.p+1. In practice, spines and dendrite surfaces
in the images are quite irregular and additional criteria must be
satisfied before a final candidate is accepted. To be accepted as a
true spine, the candidate is required to satisfy a heuristic
protrusion criterion. Let
D.sub.S.fwdarw.P.sub..sub.1.sup..sub.P.sub..sub.2.sup.C.sub.i
denote the perpendicular distance from S to the line segment 1 P 1
C i P 2 C i _
[0079] of C.sub.i and let
D.sub.E.fwdarw.P.sub..sub.1.sup..sub.P.sub..sub.- 2.sup.C.sub.i
denote the perpendicular distance from the base voxel E.sup.C.sub.i
to the line segment 2 P 1 C j P 2 C j _
[0080] as illustrated in FIG. 3(c). The spine candidate C.sub.j is
required to satisfy
D.sub.S.fwdarw.P.sub..sub.1.sup..sub.P.sub..sub.2.sup.C.sub.i.gtoreq.D.sub-
.E.fwdarw.P.sub..sub.1.sup..sub.P.sub..sub.2.sup.C.sub.i (1)
[0081] Criterion (1) clearly requires that favorable spine
candidates protrude out further from the dendrite surface than they
protrude into it. It however favors spines with narrow necks, as
the following argument shows. Consider an ideal, symmetric spine
whose base is the arc of a circle with radius R as sketched in FIG.
3(c). (Clearly i plays the integer role of the radius for the
digitized candidates C.sub.i.)
[0082] Let 2 W denote the distance between P.sub.1 and P.sub.2.
Simple trigonometry gives 3 D E P 1 P 2 = R - R 2 - W 2 , ( 2 ) D S
P 1 P 2 = R 2 - W 2 . ( 3 )
[0083] Then the protrusion criterion (1) requires
R.gtoreq.2W/{square root}{square root over (3)} in order to accept
a spine candidate, favoring narrow-necked spines and rejecting
wide-necked spines. Since R will increase faster than W for true
spines (i.e., i increases faster than the distance 4 P 1 C i P 2 C
j _ ) ,
[0084] instead of testing only the candidate C.sub.j for the
protrusion criterion, a range of candidates C.sub.i are tested for
which i is close to j. Specifically, candidates having values i in
the range
d.sub.b(S)-d.sub.p.ltoreq.i.ltoreq.d.sub.b(S)-(d.sub.p+d.sub.e1)/2,
where d.sub.e.ident.d.sub.b(E.sup.C.sub..sub.R), are considered. If
one of the candidates C.sub.i with i value in this range satisfies
(1), then the candidate C.sub.j is accepted as a true spine;
otherwise the protrusion is rejected as a spine.
[0085] The 3-D algorithm proceeds in a similar fashion except that
instead of using the projected pair of surface boundary points, the
entire ring of surface boundary points in 3-D are considered. The
minimum "surface points to backbone" distance is used to find the
reference candidate C.sub.R to correct for the local thickness of
the dendrite. The distance from S (or E) to the ring of voxels is
calculated by measuring the perpendicular distance of S (or E) to
the plane that best fits the ring of voxels in 3-D.
[0086] A.3.c Component Elimination
[0087] Spines touching the boundary of the imaged region are
ignored as they are incomplete. This is also a useful technique for
eliminating "debris" and other axons or dendrites in the background
of the image that are near or touching the dendrites of interest.
The inventive algorithms are written to allow for the imposition
one or more non-overlapping polygonal areas on the plane of the
image slices. The interior of the union of these polygons is
regarded as the region of interest for the spine detection
algorithm; any structure exterior to the polygons is ignored. By
setting the polygonal edges to cross through unwanted structures
they are also automatically ignored. As mentioned, detached
components further from the dendrite surface than a maximum
distance are also eliminated.
[0088] A.3.d Component Merging
[0089] As a spine may be identified from multiple detached "head"
and attached "base" components, a final merging algorithm which
accounts for the position and orientation of all possible spine
pieces is performed. The merging algorithm considers every
component, checking for possible merges with other components. Any
merged entity is reconsidered as a new single component, and
rechecked for possible further merges.
[0090] Merging can occur between two detached (DD) components (the
merged entity is still considered a detached component) or between
detached and attached (DA) components (the merged entity is then
considered to be an attached component). Two criteria are employed
for DD or DA type merging.
[0091] The first criterion is maximum separation; the two
components to be merged are required to be close enough (a
center-of-mass to center-of-mass separation.ltoreq.3 .mu.m). The
second criterion requires appropriate relative orientation of the
two components as demonstrated in 2-D in FIG. 4. For DA type
merging, the tip S of the attached component A is required to lie
within the triangle DP.sub.1P.sub.2. (In 3-D, the tip S is required
to lie within the cone determined by D and the ring of
spine-surface boundary points.) For DD type merging, the average
angle subtended by the center of mass of each spine with the
surface voxel locations of both spines is required to be less than
30.degree..
[0092] A.4 Image Registration and Spine Tracing
[0093] A time sequence of 3-D images must be registered to correct
for possible translational movement of the specimen. After
registration, individual spines are then traced and identified
through the image sequence.
[0094] Each consecutive pair of images F.sub.i and F.sub.i+1 are
co-registered using the spines separately identified in each image.
The offset {right arrow over (o)}=(o.sub.x, o.sub.y, o.sub.z) of
F.sub.i+1 with respect to F.sub.i is allowed to vary within a
window .vertline.o.sub.x.vertline..ltoreq.w.sub.x,
.vertline.o.sub.x.vertline..l- toreq.w.sub.x,
.vertline.o.sub.x.vertline..ltoreq.w.sub.x,.vertline.o.sub.-
y.vertline..ltoreq.w.sub.y,.vertline.o.sub.z.vertline..ltoreq.w.sub.z.
Only integer voxel offsets are considered. A conventional
registration method maximizes the cross correlation of two images;
thus no decision can be made until the correlation arrays are
computed for all offsets. Instead, an efficient sequential search
method, as disclosed by Bamea and Silverman, 1972 in "A class of
algorithms for fast image registration," IEEE Trans. Computers,
1972, vol. C-21, 179-186, is utilized which computes the l.sub.1
norm (absolute value sum) image difference
.epsilon.({right arrow over
(o)})=.SIGMA..sub.i.SIGMA..sub.j.SIGMA..sub.k.-
vertline.F.sub.i(i,j,k)-F.sub.i+1(i-o.sub.x,j-o.sub.y,k-o.sub.z).vertline.-
.
[0095] over all offsets {right arrow over (o)} in the window for
which .epsilon.({right arrow over (o)}) is less than a
predetermined threshold value T. The offset {right arrow over (o)}
with minimum .epsilon.({right arrow over (o)}) provides the optimal
registration. In practice, w.sub.x=w.sub.y=w.sub.z=5 voxels, and T
is the average number of total spine voxels in F.sub.i and
F.sub.i+1.
[0096] Individual spines are traced through the time-series. Two
spines at different times are considered to be the same if their
percentage overlap (measured in voxels) is larger than 25% of the
volume of at least one of them.
[0097] A.5 Morphological Characterization
[0098] Spine length, density and volume are computed in accordance
with one aspect of the present invention. Spines are also
classified according to their shape.
[0099] For a detached spine (without any attached component), the
spine length is determined by the distance from the recorded
dendrite surface voxel (corresponding to the associated dendrite)
to the furthest spine voxel (corresponding to the detached spine)
from the dendrite. For spines that are fully or partially attached
(consisting of a base and one or more detached components) to the
dendrite, the spine length is determined by the distance from the
center of mass of the base boundary points to the furthest spine
voxel (possibly detached from the dendrite). For the images
analyzed in the Examples herein, the automated spine length
measurement is calculated from a 2-D projection. The reason for
this is that the manual spine analysis measurements against which
the automatic analysis results are to be compared are performed in
2-D by projecting the 3-D stack of image slices along the optical
direction.
[0100] Spine density is computed as the number of spines per unit
length of dendritic backbone. For purposes of comparison with
manually analyzed images which are analyzed in 2-D projection only,
backbone length is also measured from a 2-D projection onto the
slice plane.
[0101] Spine volume is measured according to the intensity values
of the deconvolved greyscale image. For 2PLSM, the excitation of
fluorescence is limited to a sub-femtoliter focal volume
(.apprxeq.0.5.times.0.5.times.1.- 5 .mu.m.sup.3) which is larger
than that of individual spines. The intensity value recorded for
each voxel in a spine is a sum of the fluorescence from all dye
molecules excited within the focal volume. The maximum intensity
voxel near the center of a spine is therefore a measure of the
volume of a spine. As the larger cross-sectional areas of a
dendrite are typically larger than the maximum cross-sectional area
of the focal region, the maximum voxel intensity recorded along the
dendrite backbone is a measure of the size of the focal volume,
assuming the fluorescence is saturated near the center of the
dendrite. See, Svoboda et al., Science, 272, pp. 589-593 (1996) and
Sabatini and Svoboda, Nature, 408, pp. 589-593 (2000). The spine
volume is defined as the ratio of the maximum spine intensity to
the maximum dendrite intensity multiplied by an empirically
determined focal volume, 5 Spine Volume = Maximum Spine Intensity
Maximum Dendrite Intensity .times. Focal Volume .
[0102] The following classification of spine shapes is used stubby,
thin, mushroom. Spine shape is decided based on spine length (L),
head diameter (d.sub.h) and neck diameter (d.sub.n). In general
terms for thin spines, spine length should be much greater than the
neck diameter (L>>d.sub.n). For mushroom spines, spine length
should not exceed neck diameter by more than a factor of 5, and the
head diameter should be much greater than the neck diameter
(d.sub.h>>d.sub.n). For stubby spines, the neck diameter is
approximately equal to the length of the spine. The specific
criteria adopted in this classification utilize the ratios
L/d.sub.n and d.sub.h/d.sub.n to classify theirshape as summarized
in Table 1.
1TABLE 1 Ratio criteria for the classification of stubby, thin and
mushroom spines. d.sub.h/d.sub.n L/d.sub.n [0,1.3) [1.3,3)
[3,.infin.) [0,2/3) stubby mushroom mushroom [2/3,2) stubby stubby
stubby [2,3) stubby mushroom mushroom [3,5) thin mushroom mushroom
[5,.infin.) thin thin thin
[0103] A.6 Measurement of the distribution of fluorescent
proteins.
[0104] Segmentation is performed on data from one fluorescence
channel. The other fluorescence channel can be used to measure the
distribution of protein components, such as chimeric proteins
linked to GFP. The analysis is simply to measure the fluorescence
on the second channel as a measure of protein concentration, in
pixels that were previously determined to belong to the neuronal
structure.
[0105] B. Image Acquisition
[0106] From a data analysis standpoint CLSM and 2PLSM provide
essentially equivalent challenges. However, to gain an
understanding of the dynamics of neuronal circuits, neurons should
preferably be studied in preparations that are as intact as
possible. For many questions of sub-cellular physiology, as stated
above, the living brain slice offers an attractive compromise
between the obvious limitations of cultured dissociated neurons and
the experimental difficulties encountered when working with intact
animals.
[0107] One problem with brain slice physiology has been that
scattering of light makes traditional optical microscopies,
including CLSM, difficult in living tissues. For these reasons, as
also stated above, the data is preferably collected using 2PLSM,
which allows high resolution fluorescence imaging in brain slices
up to several hundred microns deep with minimal photodamage.
[0108] At present, parameters that require routine adjustments
include the region of interest and segmentation threshold. Other
parameters that are used in the deconvolution, backbone extraction,
spine component elimination and tracing algorithms are empirically
determined; they remained the same throughout all of the examples
that have been described herein. Segmentation is crucial to the
analysis due to the relatively low intensity associated with small
spines and the high intensity of the dendrites. Choosing a critical
threshold is important; simple thresholding is adequate for most
images that have been preprocessed by median filtering and
deconvolution.
[0109] It is contemplated to provide the algorithm of the present
invention within a server accessible via a network, such as the
Internet or a local area network (LAN), by a plurality of users.
The users can then transmit data to the server for processing using
the algorithm. The results would then be transmitted by the server
to the user via e-mail or by other known techniques.
[0110] The following examples are included for purposes of
illustrating certain aspects of the present invention and are not
intended to limit the invention as defined by the claims
herein.
EXAMPLE 1
Sample Preparation and Microscopy
[0111] Cultured hippocampal brain slices were prepared from 7 day
old rats. After five days in vitro a small subset of neurons were
biollistically transfected (for example, Lo, et al., (1994).
Neuronal transfection in brain slices using particle-mediated gene
transfer, Neuron 13, 1263-1268, the contents of which are hereby
incorporated by reference) with a plasmid carrying the gene for
enhanced green fluorescent protein (GFP) (commercially available
from Clonetech) At least two days after transfection, slices were
transferred to a perfusion chamber for imaging. Labeled neurons
were identified and imaged using a custom-made 2PLSM laser scanning
microscope (as described in Mainen, Z. F., Maletic-Savatic, M.,
Shi, S. H., Hayashi, Y., Malinow, R., and Svoboda, K. (1999),
Two-photon imaging in living brain slices, Methods 18, 231-239, the
contents of which are hereby incorporated by reference). The light
source was a Ti:sapphire laser running at a wavelength of
.apprxeq.990 nm (repetition frequency 80 MHz; pulse length 150 fs).
The average power delivered to the backfocal plane of the objective
(40x, NA 0.8) varied depending on the imaging depth (range 30 to
150 mW). Fluorescence was detected in whole-field detection mode
with a photomultiplier tube.
EXAMPLE 2
Static Analysis
[0112] To validate the automatic spine detection algorithm, an
experiment, E.sub.1, involving .apprxeq.200 spine measurements over
15 dendrites of hippocampal CA1 and a small number of CA3 neurons
was performed. The same imaged regions were subjected to both
automatic and manual analysis. A total of 174 spines were
identified by both methods; an additional 10 spines were identified
only by the manual method; and a further 28 spines were identified
only by the automatic method.
[0113] The results of the manual and automated analysis for one of
the dendrites in this experiment are illustrated in FIG. 5(a)-(b).
Twenty-one spines were detected by both methods; three additional,
relatively short, spines were detected by the automatic method.
FIG. 6 compares the individual spine lengths, average spine length
and spine density measured for this particular dendrite. Spines 8
and 22 demonstrate the difficulties encountered by both detection
methods when two spines appear to overlap. For spine 8, the manual
detection has identified only the shorter of two spines which
appear to overlap; whereas the automatic method has identified only
the longer. For spine 22, again two spines appear to overlap and
are considered as a single spine by the automatic method. On the
other hand, the manual method failed to identify either of
them.
2TABLE 2 Measured mean spine lengths (.+-. standard deviation) for
the spines in experiment E.sub.1. Mean spine length Population size
Method (.mu.m) 174 Manual 1.05 .+-. 0.62 174 Automatic 1.08 .+-.
0.63 10 Manual 0.75 .+-. 0.57 28 Automatic 0.38 .+-. 0.28
[0114] Table 2 compares the mean spine length measured by each
method for the population of 174 spines detected in common. The
mean lengths for those spine detected by only one of the methods
are also presented. For the common detected spine population, the
manual and automatic spine length measurements agree to within one
standard deviation, though the standard deviations are large. (The
large standard deviation is partly due to averaging over spines of
different shape classification.) A paired samples Student's t-test
to determine whether the difference in measurements by the two
methods is significant provides a stronger test of the agreement
between the two methods of length measurement. Column 1 of Table 3
summarizes the results of the paired t-test; there is no
significant difference between the two methods of length
measurement.
3TABLE 3 Paired samples t-test results for measured spine lengths
and densities, experiment E.sub.1. Spine length Spine density
Degrees of freedom 173 14 t statistic -1.24 -0.87 p-value 0.22,
two-sided 0.40, two-sided
[0115] A one-way ANOVA was used to test for any dependence of
dendrite origin on the observed differences in measured spine
length. The test produces an F statistic value of 0.95
(d.sub.num=14 and d.sub.den=159) with a p-value of 0.51. Thus the
differences in spine length measurements between the two methods
are uniform across the different dendrites.
[0116] The mean spine length for the 28 spines detected only by the
automatic method of the present invention indicates a population of
smaller length spines. The results in Table 4 of an independent
samples t-test for these 28 spines shows that the difference in
spine lengths is significant compared to that obtained by either
measurement method for the population of 174 commonly detected
spines. The results indicate that the automatic algorithm of the
present invention is detecting short spines more consistently than
the manual method.
4TABLE 4 Independent samples t-test results, experiment E.sub.1.
Method (population Manual (n = 174) Automatic (n = 174) Automatic
only df = 200 df = 2 (n = 28) t = 5.63 t = 5.82 p = 6 .times.
10.sup.-8, two-sided p = 2 .times. 10.sup.-8, two-sided Manual only
df = 182 df = 182 (n = 10) t = 1.51 t = 1.65 p = 0.13, two-sided p
= 0.10, two-sided
[0117] For the 10 spines detected only by the manual method, the
mean length measurement lies midway between that obtained for the
common and automatic only populations. An independent samples
t-test (Table 4) shows no significant differences with the
measurements obtained by either method for the 174 commonly
detected spines. The results (df=36, t=2.65, p=0.01, 2-sided) of an
independent samples t-test between these 10 spines and the 28
detected only by the automatic algorithm indicate a significant
difference between the spine lengths of these two populations.
[0118] Visual observation of these 10 spines reveals that 7 of them
touched the boundary of the image region and were consequently
rejected by the automatic algorithm. The remaining 3 were not
resolved by the automatic algorithm as each touched some
neighboring spine (which was detected). This is thus a reflection
of the combined effectiveness of the median filter, deconvolution
and simple thresholding algorithms in segmenting the images. Based
upon visual investigation of the images, no more than 2% of the
spines were estimated not to have been resolved by the automatic
algorithm due to segmentation related effects.
5TABLE 5 Measured mean spine density (.+-. standard deviation for
the dendrites in experiment E.sub.1. Mean spine density Population
size Method (.mu.m.sup.-1) 15 Manual 0.45 .+-. 0.09 15 Automatic
0.47 .+-. 0.15
[0119] Table 5 compares the mean spine densities separately
measured by each method for the 15 dendrites. (For the manual
method this is a total population of 184 spines; for the automatic
method, 202 spines.) The density measurements by either method
agree to within one standard deviation. For the paired sample of 15
dendrites, a Kolmogorov-Smimov test shows that the dendrite by
dendrite difference in the automatic and manual measured densities
is very close to normal, so that a paired-dendrite samples t-test
can be applied. Column 2 of Table 3 summarizes the result; the
automatic spine density measurement is not significantly different
from the manual.
EXAMPLE 3
Static Analysis
[0120] For spine volume measurement and shape classification,
automated results are reported herein as no manual determination is
available. A second experiment, E.sub.2, was performed under the
same experimental conditions to increase the sample size
(E.sub.1+E.sub.2) up to 700 spines. The spine volumes were
calculated from the ratio of the maximum intensity values of the
spine to the dendrite, as described in Section A.5 using an
empirically determined focal volume of 0.5.times.0.5.times.1.5
.mu.m.sup.3. FIG. 7 shows the volume-length correlation plot for
the spines measured in experiments E.sub.1 and E.sub.2 according to
their determined classification. The mushroom shaped spines occupy
the widest spectrum of lengths and volumes. The ratio of
stubby:mushroom:thin spines is 0.54:0.36:0.10.
6TABLE 6 Comparison of automated spine length and volume
measurements of hippocampus cells in PND 7 cultured neurons with
the PND 15 SSEM results of Harris et al., "Three-dimensional
structure of dendritic spines and synapses in rat hippocampus (CA1)
as postnatal day 15 and adult ages: implications for the maturation
of synaptic physiology and long-term potentiation," J. Neurosci.
1992, vol. 12, 2685-2705. Shape classification Measurement Method
stubby mushroom thin volume (.mu.m.sup.3) automatic 0.07 .+-. 0.04
0.06 .+-. 0.05 0.06 .+-. 0.04 SSEM 0.11 .+-. 0.07 0.18 .+-. 0.09
0.05 .+-. 0.03 length (.mu.m) automatic 0.65 .+-. 0.37 1.35 .+-.
0.55 1.38 .+-. 0.54 SSEM 0.65 .+-. 0.38 0.95 .+-. 0.30 1.40 .+-.
0.39
[0121] Table 6 summarizes the average spine volume and length
measurements in each shape category and presents comparison with
the SSEM results (Harris et al. 1992; Table 4) on rat hippocampus
CA1 cells for postnatal day (PND) 15 animals. All automatic
measurements are within 1.5 standard deviations of the SSEM result,
though the volume results are generally smaller. It is noted
however that the automatic results come from cultured neurons and
younger aged animals. In addition, no corrections for any
fixation-induced changes have been performed in the SSEM study.
EXAMPLE 4
Time series analysis
[0122] Time-series data provides the ability to capture dynamic
changes in dendritic spine morphology. A series of 50 3-D images
was taken at 30 second intervals spanning a time period of 25
minutes. FIG. 8(a) shows the number of spines detected in the
images as a function of time using the automatic method. On average
27.5.+-.2.3 spines were detected in each image. An interest lies in
the question of the frequency of observation of any particular
spine over time.
[0123] In total, 52 spines were detected and traced through the
time-series. FIG. 8(b) shows how the observations of the 52 spines
were distributed in time. The spines were indexed (1.fwdarw.52)
according to the first time in which they appeared. Thus, 27 spines
were observed over the full 25 minutes of image taking; among
those, 16 were present at all time points.
[0124] FIG. 9(a) summarizes the distribution of spine length and
volume as a function of time. Both length and volume distributions
are skewed, with smaller lengths and volumes dominating, consistent
with the stubby, mushroom, and thin spine ratios noted above. The
dynamics of the spines are measured using an index for spine
motility, which is defined as the summed difference in length of a
spine in time divided by the total number of time steps. FIG. 9(b)
plots the distribution of motilities for the 52 spines traced in
this series. For this limited data set, the number of spines (n)
decreases with motility (m) approximately as
n(m)=n(0)e.sup.-3.69m.
[0125] A comparison between automated and manual length
measurements was made for a limited subset of this time series
data; manual length measurements were made on a subset of 10 of the
52 spines. FIG. 10 presents comparisons between the automated and
manual spine length measurements as a function of time for 5 of the
spines chosen to represent different average lengths. Consistently
longer lengths were measured by the manual method for the longer
spines (1 and 2). For the medium length spines (3 and 4), the
manual and automated results are very similar. For the short spine
(5), some deviations are observed; occasionally the spines were not
detected by the automatic method.
[0126] A paired t-test was performed on this subset of 10 spines to
determine whether the average spine length determined by the
automated measurement is significantly different from that
determined by the manual measurement. A Kolmogorov-Smimov test
shows that this sample of 10 difference measurements (-0.32.+-.0.17
.mu.m) is very close to normal so that a paired samples t-test can
be applied. The observed t statistic value is -2.79, with p=0.02,
two-sided, revealing some significance in the averaged
difference.
[0127] While this is contrary to the results in experiment E.sub.1,
it is noted that the manual measurements were made by a different
user than in E.sub.1. Pearson correlation was therefore used to
test whether the automatic and manual measurements are correlated
in time. The correlation values (r) obtained for the 10 spines for
n=50 time points range from 0.92 to 0.29; with two-sided p-values
ranging from 0.00 to 0.04. Thus, for these 10 spines significant
correlations between the automatic measurements and the manual
measurements in time are observed. Therefore, the existence of a
systematic bias between the manual and automated measurements for
this set of data that can be attributed to a change in the user
making the manual measurements. The significant Pearson correlation
however indicates that the manual measurements are duplicating the
trends found by the automatic measurements.
EXAMPLE 5
Evaluation of the Effects of mTOR Kinase on Spine Formation
[0128] mTOR kinase controls the phosphorylation of the translation
regulators p70.sup.56k and 4E-BP1. It is a central regulator of
cell growth and highly expressed in dendrites and highly associated
with synaptic proteins. Rat hippocampal CA1 neurons were
transfected by biollistic gene transfer with GFP alone or together
with the wild-type kinase (mTOR wt) or a non-functional mutant
(mTOR kd) that acts as a dominant negative (for details of plasmids
see: Sabatini et al., (1999). Interaction of RAFT1 with gephyrin
required for rapamycin-sensitive signaling, Science 284, 1161-4,
the contents of which are hereby incorporated by reference) (for
methods see Example 1). The 3d structure was imaged using 2PLSM as
described in Example 1 and the inventive algorithm utilized to
analyze dendritric shape. As can be seen in FIGS. 11(A) and 11(B)
mTOR kinase controls spine size and density, i.e., mTOR wt
increased spine density and mTOR kd decreased spine density as
compared to the GFP control. Data from at least 5 neurons were in
every group.
EXAMPLE 6
Evaluation of the Effects of Overexpression of Neuroligin on Spine
Formation
[0129] Neuroligin (NLG) is believed to be involved in synapse
formation. When nonneuronal cells are engineered and cultured to
express NLG, they are associated with the development of
presynaptic structures in contacting axons, suggesting that
NLG-neurexin interactions are a key step in synapse formation. Rat
hippocampal CA1 neurons were transfected by biollistic gene
transfer with GFP alone or together with NLG or a mutant form of
NLG that does not contain the AchE domain (designated "AChE") (See,
Scheiffele et al., (2000). Neuroligin expressed in nonneuronal
cells triggers presynaptic development in contacting axons, Cell
101, 657-69, the contents of which are hereby incorporated by
reference). The 3d structure was imaged using 2PLSM as described in
Example 1 and the inventive algorithm utilized to analyze dendritic
shape. As can be seen in FIGS. 12(A) and 12(B), overexpression of
NLG did not produce a detectable phenotype change in spine density
or spine shape. Data from at least 5 neurons are in every
group.
EXAMPLE 7
Measurement of the distribution of GluR1-GFP in dendritic
spines.
[0130] Glur1 is a synaptic receptor that is thought to play an
important role in synaptic plasticity. To measure the distribution
of Glur1 in dendritic spines neurons were transfected with a virus
expressing Glur1-GFP (for methods see Shi, S. H., Hayashi, Y.,
Petralia, R. S., Zaman, S. H., Wenthold, R. J., Svoboda, K., and
Malinow, R. (1999). Rapid Spine Delivery and Redistribution of AMPA
Receptors After Synaptic NMDA Receptor Activation, Science 284,
1811-1816, the contents of which are hereby incorporated by
reference). Neurons expressing glur1-GFP were then patch-clamped
and filled with a red fluorohore (Texas Red, Molecular Probes). A
two color image was then acquired using 2PLSM, with an excitation
wavelength of 910 nm. The red image was used for segmentation and
to perform a spine analysis. The green image was used to estimate
the distribution of Glur1-GFP in dendrites and spines.
[0131] It will be understood that various modifications may be made
to the embodiments and examples disclosed herein. For example,
alternative algorithms can be created to accomplish the criteria
set forth above. It should be understood that, notwithstanding the
emphasis on spine morphology, the algorithms described herein can
be applied to the larger axonic, dendritic and cell body structures
of neurons to determine length, volume, shape classification and
density. Therefore, the above description should not be viewed as
limiting, but merely as exemplifications of preferred embodiments.
Those skilled in the art will envision other modifications within
the scope and spirit of the claims appended hereto.
* * * * *